Content uploaded by Sumair Aziz
Author content
All content in this area was uploaded by Sumair Aziz on Dec 15, 2023
Content may be subject to copyright.
978-1-6654-6472-7/23/$31.00 ©2023 IEEE
An Automated System for the Classification of
Bronchiolitis and Bronchiectasis Diseases using
Lung Sound Analysis
Syed Ali Faraz Jaffery, Sumair Aziz, Muhammad Umar Khan*, Syed Zohaib Hussain Naqvi, Muhammad Faraz
and Adil Usman
Department of Electronics Engineering,University of Engineering and Technology Taxila,Pakistan
*email: sa.umarkhan@gmail.com
Abstract— The main goal of this paper is to develop a
classification model and a technique to identify bronchiolitis and
bronchiectasis using lung sound analysis. In this paper, we
develop a methodology to automatically identify lung disease
through an intelligent system. ICBHI lungs sound database was
used for this study. A total of 64 lung recordings, selected from
three pulmonary classes namely normal, bronchiectasis and
bronchiolitis were used for this purpose. To accomplish the task,
we first split all the recorded signals into four parts to increase
the number of input data. Discrete wavelet transform was used
to denoise and segment the pulmonological data. Mel frequency
cepstral coefficients were then computed from the cleaned
signal. After extensive experimentation with various classifiers,
the highest recognition rate of 99.6% was found by using K-
Nearest Neighbors.
Keywords; Bronchiolitis, Bronchiectasis, lung sounds,
Discrete Wavelet Transform, MFCC, K-Nearest Neighbors,
Feature extraction
I. INTRODUCTION
Over the past few years, pulmonary abnormalities have
become a common issue all over the world. Chronic
obstructive pulmonary disease (COPD), asthma,
bronchiectasis, bronchiolitis, upper respiratory tract infection
(URTI), pneumonia, lower respiratory tract infection (LRTI)
and other diseases are some major respiratory diseases.
Among them, the mortality rates for bronchiolitis and
bronchiectasis are frequently underestimated due to
differences in diagnostic criteria. In the U.S., 110,000 infants
are hospitalized every year due to bronchiolitis [1]. Typically,
bronchiolitis is caused by a viral infection, which, in most
cases, is caused by the respiratory syncytial virus. RSV virus
is quite prevalent and is easily spread by coughs and sneezes.
Almost all infants have it by the age of two [2]. Infants with
bronchiolitis have damage in the small airways that can cause
coughing, wheezing, and breathing difficulties.
In another study by the Australian Institute of Health and
Welfare (AIHW), there were 983 deaths reported in 2018 that
had bronchiectasis described either as the underlying cause of
death (387) or as an associated cause (596) of death [3].
Bronchiectasis is a chronic disease that causes the abnormal
widening of the lungs' airways. Because of these harmed
airways, mucus and stones gather in the lungs. It causes lung
damage and prevents healthy airway function by blocking
them with mucus [4]. Bronchiectasis can also result in
wheezing cough pain, joint pain, and shortness of breath, and
all these symptoms can also be caused by bronchiolitis. But,
bronchiectasis is induced by pneumonia, pertussis,
tuberculosis, and nontuberculous mycobacterial infections
that lead to abnormal windings in the lungs of older
individuals. In contrast, bronchiolitis is an acute lower
respiratory infection that is commonly induced by viral
infections in infants [5]. The major difference between these
two diseases is that bronchiectasis is a chronic condition that
mostly affects older people. However, bronchiolitis typically
affects children and can be treated at home.
The clinical options for identifying respiratory diseases are
many. Imaging techniques including magnetic resonance
imaging (MRI), computed tomography (CT) scans, and chest
X-rays are used to diagnose respiratory illnesses. Contrarily,
adopting such imaging modalities presents several difficulties,
particularly for many patients in third-world countries. These
difficulties include the risk of receiving repeated dosages of
hazardous rays, the expense of the equipment, and the
inconvenience of deploying it in remote areas [6]. Despite
faster and continuous technological advances in the diagnosis
of chest diseases, auscultation keeps the most extensively used
and effective lung disease diagnostic tool [7]. Respiratory
sounds captured using a stethoscope are a direct indicator of
lung health and abnormalities. They yield essential details and
significant information regarding the condition of the lungs.
The use of a stethoscope provides a restricted and biased
perspective of breathing sounds. Subjectivity causes
differences in the perception of lung sounds (LS) by different
medical specialists. Subjectivity and discrepancy are made by
the pathologist's ability to hear, experience, discriminate and
define various sound patterns. Moreover, the stethoscope
output is sensitive to noise. Such noisy signals obscure critical
aspects of LS signals and result in inaccurate lung illness
diagnoses [8].
Today's era of computing technology has achieved
significant strides in the early and quick identification of a
variety of respiratory disorders. To overcome the limitations
of physical diagnosis, various machine/deep learning
algorithms for automated lung disease recognition based on
lung sound data have been presented in the literature.
Moreover, most of these pulmonary disease classification
approaches are developed for a particular and limited lung
disease classification that is less precise, and ineffective for
the evaluation of non-stationary LS signals. However, this
study describes the development of an automated and
intelligent system for the accurate detection and classification
of two lung disorders, bronchiolitis, and bronchiectasis, using
signal processing and machine learning algorithms.
2023 International Conference on Robotics and Automation in Industry (ICRAI) | 978-1-6654-6472-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICRAI57502.2023.10089608
Authorized licensed use limited to: University of Canberra. Downloaded on June 10,2023 at 14:01:12 UTC from IEEE Xplore. Restrictions apply.
II. RELATED WORK
Currently, a machine learning model has been developed
to identify a single pulmonary pathology from the analysis of
lung sounds. [9]. A timely and precise diagnosis can reduce
the risk of mortality. However, the subjective nature of
abnormal noises like coughs has made the identification of
pneumonia, bronchiolitis, and other lung conditions extremely
challenging[10]. In the study conducted by Z. Tariq et al. [11],
the ICBHI dataset was used and six pulmonary diseases are
classified by using a convolution neural network (CNN), and
the feature vector is computed from the Spectrogram of
annotated lung sound samples. The classification accuracy
was found to be 97%. Garcia-Ordas et. al [12] presented a
convolution neural network (CNN) with variational
autoencoders (VAE) to classify lung disorders. The audio
signals were represented by Mel spectrograms. By using the
CNN model, they identify six distinct pathological diseases.
This study achieved 99.1% sensitivity and 99.4% specificity.
V. Base et al. in [13] purposed the deep neural network model
for the identification of the LS. These researchers have used
the MFCC features with the RNN classifier and achieved
95.67% accuracy on the six pulmonary diseases classification.
Another study conducted by [14] S. I. Khan et al. computed
tomparal and frequency domain features from the first four
IMFs of EMD and these IMFs were then further processed for
the two dimensions and higher dimensions space
representation. These researchers achieved the highest
performance of the ensembled bagged tree classifier with an
accuracy of 96.29% on 10-fold cross-validation. In [15] M.
Orders et al. implement the convolution neural network
(CNN) to identify pulmonary abnormalities. They used the
Mel Spectrogram for a visual representation of lung sounds.
The recognition rate for the six different classes was achieved
by 99%. In another research [16] L. Brumes et al. purposed a
neural network-based model for the identification of
pulmonary diseases. The feature vector consists of the
combination of time, cepstral and spectral for the input of the
classification learner. The model achieved an F-score of
98.3%. In [17] L. Fragipan et al. derived a novel method to
recognize the multi-class classification of pulmonary illness
from LS analysis. They used a boosted decision trees classifier
with Shannon entropy, log entropy, and spectral entropy
features to classify the pulmonary illness and achieved
98.27% accuracy. In [18] J. Acharya et. al presented a hybrid
CNN-RNN learning model consisting of three stages namely
deep CNN, Bi-LSTM, and SoftMax layers for the
classification of pulmonary diseases. These researchers also
implement a subject-specific classification model in which the
last 3rd stage of the proposed hybrid CNN-RNN is updated
according to the patient-specific data. Moreover, for the
reduction of memory cost, a weight quantization technique is
also discussed in the same study in which the amplitude of
every layer's weight is quantified in the log domain. Such
models are difficult to train due to the challenges of collecting
a vast amount of subject-specific data. While a subject-
specific model demands more time and effort from medical
experts for gathering and grouping the data. Previous studies
for computerized-based respiratory LS detection have been
conducted by using the application of signal processing and
machine learning ML algorithms to automate the diagnosing
mechanism however a great deal of work is still required in
this research area. A complete comparison of previous studies
is given in Table I. CNN requires a large dataset in order to
attain better results. The dataset in the case of lung sound is
limited. CNN also needs more computing resources, so can
not be deployed on low-cost embedding devices which are
battery-powered. There is a substantial need for an accurate
TABLE I. COMPARISON WITH PREVIOUS STUDIES
Study Database
Feature
extraction
Classification
Results
[11]
ICBHI
2017
Spectrogram
CNN
Sensitivity (97%)
[12]
ICBHI
2017
Mel
-
Spectrogram
CNN
99.1
% (
for
muti
-
class classification)
98.8% (for binary classification)
[13]
ICBHI
2017
MFCC Features
RNN
Sensitivity (95.67%)
F1-Score (95.66%)
Kappa (94.74%)
[14]
ICBHI
2017
F
our IMF features
calculated from EMD
Ensemble of
b
agged tree
Classifier
Sensitivity (9
6.29
%)
[15]
ICBHI
2017
Mel
-
Spectrogram
CNN
Sensitivity (99.00%)
F1-Score (90.00%)
[16]
ICBHI
2017
Time, cepstral and spectral
feature from
recorded sound
N
eural
N
etwork
Algorithm
F
-
score
(
98.3%
)
[17]
ICBHI
2017 +
Self-
Recorded
Entropy features
Boosted DT
Sensitivity
(98.27%)
Specificity (95.28%)
F1-score (98.9%)
Kappa (92.28%)
[18]
ICBHI
2017
Mel
-
Spectrogram
Hybrid CNN
-
RNN
+
Patient Specific Model
(
66.31%
) for hybrid model
(71.81%) for patient specific Model
Authorized licensed use limited to: University of Canberra. Downloaded on June 10,2023 at 14:01:12 UTC from IEEE Xplore. Restrictions apply.
lung sound analysis based on pattern recognition and machine
learning.
III. PROPOSED METHODOLOGY
Figure 1 illustrates the proposed methodology of our
intelligent system for the detection of respiratory lung sound
LS diseases.
Firstly, bronchiolitis, bronchiectasis, and normal sounds
are taken from the Lung sound LS database and split each
recording into four parts to increase the total number of data
samples. In the second stage, pre-processing of these signals
is performed by using DWT. In the next stage, MFCCs
features are extracted and then classification is performed on
selected features by using the KNN classifier.
A. Lungs sound database
The publicly available dataset used in this research is
ICBHI 2017 which contains lung sound files as shown in
Figure 2. The human lung sounds in the ICBHI database were
recorded over several years by two different research teams.
The dataset contains a 5.5-hour audio recording of 126
subjects' lung sounds which are recorded with four types of
recording equipment. The lung sound recording contains the
6898 respiratory cycles that include the crackles, wheezes, a
combination of both, and annotated audio samples. The
dataset includes normal, bronchiectasis, bronchiolitis,
pneumonia, asthma, COPD, URTI, and LRTI, which describe
different lung conditions. Among these sounds, we used 64
audios in this research, which are normal, bronchiectasis, and
bronchiolitis. To increase the number of audio
samples, we divided the selected samples into four parts.
Table II provides details about the dataset used in this work.
B. Preprocessing
Preprocssing is very important step in signal processing
pipeline. For feature extraction purposes, it is preffered to
have clear audio samples of lungs. There are certain sounds
like heart sounds, background sounds, and vocalization that
make complexity in the identification of certain lung diseases
because inappropriate features are measured due to these
strange sounds. These noise elements hides little differences
that exists between sounds of different categories, and thus
make a classification task very complicated. To avoid
complication, it is always preffered to remove the unwanted
components and nosie from signal before further processing.
In this study, we applied discrete wavelet transform on all the
signals for preprocessing puspose [19-22]. Figures 3 and 4
represent the time domain of respiratory signals with and
without removing unwanted sounds. It was observed that after
preprocessing the lung sound signals, high freuqney noise was
significantly reduced. Frequency analysis was also carried out
to make sure that the preprocessed signal contains the relevant
frequency bands of lung sounds.
C. Feature Extraction
Features are the most important and essential components
that are fed to the classifier to differentiate between different
classes [23, 24]. Features represent the input signal data into
numeric components, thus reducing the data dimensionality,
and removing redundant information. An appropriate feature
vector must contain the relevant data and remove unnecessary
information. There are two main properties of strong features:
1) the Mean of features within the same class must be having
a minimum variance, and 2) Centroids of features of different
classes must be having significant distance. We use MFCC
information of denoised audio signals for classification in this
research. Mel frequency cepstral coefficients [25-27] is
computed from the discrete cosine transform of a log power
spectrum on a Mel frequency scale as described in equation 1
below.
(1)
where f is equal to the frequency in hertz.
Fig. 2. ICBHI respiratory database of lungs sound
TABLE II. DATASET DETAILS
ID. Name of
Class
Audio
Files
Segmented Audio
Files
1. Healthy 35 140
2. Bronchiolitis 13 52
3. Bronchiectasis 16 64
Total 64 256
Fig.1. Proposed Method for Classification of bronchiolitis and bronchiectasis
Authorized licensed use limited to: University of Canberra. Downloaded on June 10,2023 at 14:01:12 UTC from IEEE Xplore. Restrictions apply.
To determine the MFCC coefficient from the audio input,
there are five steps to follow:
1. Frame the input wave with fixed-size windows.
2. Calculate the FFT of each frame
3. Use equation 1 to obtain the Mel spectrogram of each
segment.
4. Apply the logarithm to each Mel spectrogram to get the
log power spectrum.
5. Calculate the discrete cosine transform (DCT) of each
logarithmic power spectrum to get MFCC coefficients.
As a result, we obtain thirteen MFCC features/coefficients.
These features were further analyzed to identify best
performing parameters using back elimination method.
D. Classification
The final stage of this research is classification. After
extracting the MFCC features our audio signals pattern is
ready for classification. For the KNN classifier to recognize
the pattern of the different sounds, ten-fold cross-validations
were carried out to make sure random factors don't enter into
the equation [28-30]. The K-Nearest Neighbor algorithm [31]
is an instance-based learning model that classifies the objects
based on their nearest training instances in the feature. It used
a supervised learning algorithm to solve both regression and
classification problems. To classify a problem a KNN
algorithm requires only two parameters a distance feature and
the value of the K. The distance function for continuous
variables is calculated as follows
(2)
!"#$ % &
(3)
#'()'# * # +
,+
- (4)
And for the categorical variables, the distance function is
calculated as follows
./
(5)
The most effective and strong technique for categorization
and prediction is the decision tree. The decision tree
classification algorithm is constructed using nodes, branches,
and leaf nodes. Fine Tree is a version of the decision tree, with
a huge number of branches and decision nodes. SVM is a
widely used classification method based. SVM differentiates
the input features using a hyperplane based on support vectors.
SVM classifies complex data patterns using nonlinear kernels
such as the Cubic operator. The kernel trick is used to increase
the dimension of input data. The extended dimensions are
employed to make a better separation between the two classes.
For multiclass problems, SVM adopts one-vs-one or one-vs-
rest approaches. Multiple models are trained and results are
predicted based on voting strategy. Boosted Tree is a kind of
ensemble method, which is a combination of many weak
learners. The prediction performance is enhanced in the
ensemble method due to the combination of different
classifiers. The performance of the proposed scheme was also
tested using kernel naive Bayes classifier, which is easy to
develop but required the input features to be independent.
Quadratic discriminant analysis (QD) is another type of
classifier that requires the input features to follow Gaussian
distribution. QD is also widely employed classifier in
biomedical domain.
IV. RESULTS AND DISCUSSION
The discrete wavelet transform algorithm is applied after
augmenting the three types of lung sound data: healthy,
bronchiolitis, and bronchiectasis. It split the signals into
detailed and approximate coefficients and removes the
approximate coefficient containing lower frequencies. We
analyze that our region of interest (ROI) lies in the
approximated coefficient of DWT due to the presence of the
desired frequency of LS in it. MFCCs, which are calculated
from the detail coefficient, are mostly used to classify LS
illnesses using different classifier models. The back-
elimination method is performed for the MFCCs features
fusion in this research and after all, we get the maximum
accuracy at 8 selected MFCC features out of 13. The extracted
features were tested with a several classification methods such
as Fine Tree (FT), Fine KNN, Cubic Support Vector Machine
(SVM), Boosted Trees (BT), Kernel Naïve Bayes (KNB), and
Quadratic Discriminant Classifier (QDC). The classification
Fig. 3. Raw LS for healthy, bronchiolitis, and bronchiectasis
Fig. 4. Preprocessed LS for healthy, bronchiolitis, and
bronchiectasis
Authorized licensed use limited to: University of Canberra. Downloaded on June 10,2023 at 14:01:12 UTC from IEEE Xplore. Restrictions apply.
performance was evaluated using 10-fold cross-validation. In
10-fold cross-validation, the dataset is divided into ten parts.
In each iteration, only one part is used as a "test set", while the
remaining nine parts are combined to create a "training set".
In the next iteration, the second part is used as a test set, and
all other parts are merged to construct a training set. This
process is repeated ten times, and ten models are evaluated.
The results of all iterations are averaged to get the final
performance measures. 10-fold cross-validation is a
significant scheme for smaller datasets.
The performance in terms of accuracy for differentiating
different LS features is illustrated in Figure 5. All
classification algorithms show comparable performance,
which confirms that processing and extracted features have
strong discrimination ability. Accuracies of 85.2%, 88.7%,
82%, and 87.9% were obtained with FT, BT, KNB, and QDC,
respectively. The best results of 99.3% were obtained using
Fine KNN, followed by the second best (94.1%) obtained
through Cubic SVM.
The experimental process was further extended by
exploring different versions of KNN. Figure 6 illustrated the
results using Fine KNN, Medium KNN, Coarse KNN, Cubic
KNN, Cosine KNN, and Weighted KNN. The lowest
performance of 62.2% was provided by the Coarse KNN.
Medium KNN predicts the class of input features by
comparing the results of the 10 nearest neighbors. Similar
results of 84.4% accuracy were obtained via Medium KNN,
Cosine KNN, and Cubic KNN. The best performance of
99.3% was yielded by Fine KNN for differentiating different
lung sounds.
The confusion matrix presented in Figure 7 shows
classwise results of classification using Fine KNN. Out of 140
Healthy sounds, only 1 was mispredicted as Bronchiolitis.
Similarly, only one sound out of 64 was mispredicted in the
case of Bronchiectasis. All 52 signals of Bronchiolitis were
correctly classified.
V. C
ONCLUSION AND FEATURE WORK
In this paper we propose a novel method to detect
bronchiolitis and bronchiectasis pathologies in LS. In the
studies of LS data, it is very usual to have a less amount of
data. One of the limitations that we are challenging during this
work is the lack of data, so we decided to split every audio
sample into four parts. For denoising the audio samples from
irrelevant frequencies, discrete wavelet transformations are
applied. The DWT yields approximate and detailed
coefficients that contain un-noised information. The
approximate coefficient contains the low frequencies part of
the input signal while the detailed coefficient contains the
higher frequencies part of the input signal. We extracted
MFCCs feature from the approximate coefficient of DWT for
the input of training models. We tested these features with
classification models and found that the fine KNN classifier
achieved the best results with 99.6% accuracy on selected
features. With this promising result, we conclude that the Fine
KNN machine learning model with the input of selected
MFCC features that extract from the approximate coefficient
of DWT has demonstrated excellent performance in the
classification of bronchiolitis and bronchiectasis illnesses.
Future works could include new feature exploration,
change in classification methods, and then integration of the
system on hardware. This research work can also be enhanced
for the diagnosing of other respiratory lung diseases by using
the same methodology.
R
EFERENCES
[1] T. Ooka et al., "Proteomics endotyping of infants with severe
bronchiolitis and risk of childhood asthma," Allergy, 10/23/2022.
[2] (19 october 2022). Bronchiolitis. Available:
https://www.nhs.uk/conditions/bronchiolitis/#:~:text=Bronchiolit
is%20is%20caused%20by%20a,children%20it%20can%20cause
%20bronchiolitis.
[3] Australian Institute of Health and Welfare 2020. Bronchiectasis.
Cat. no. ACM 36. Canberra: AIHW. Viewed 07 December 2020,
https://www.aihw.gov.au/reports/chronic-respiratory-
conditions/bronchiectasis
[4] A. E. O’Donnell, "Bronchiectasis—A Clinical Review," New
England Journal of Medicine, vol. 387, no. 6, pp. 533-545, 2022.
[5] S. R. Dalziel et al., "Bronchiolitis," The Lancet, 2022.
[6] B. A. Tessema, H. D. Nemomssa, and G. L. Simegn, "Acquisition
and Classification of Lung Sounds for Improving the Efficacy of
Auscultation Diagnosis of Pulmonary Diseases," Medical Devices
(Auckland, NZ), vol. 15, p. 89, 2022.
Fig. 5. Accuracies of various classifiers
85.20% 99.30% 94.10% 88.70% 82.00%
87.90%
Fine Tree
Fine KNN
Cubic SVM
Boosted Trees
Kernel Naive Bayes
Quadratic
Discriminant
Fig. 7. Confusion Matrix of the proposed system
Fig. 6. Accuracies of different KNN Kernels
99.30% 84.40%
60.20%
84.40% 84.40% 84.40%
Fine KNN
Medium KNN
Coarse KNN
Cosine KNN
Cubic KNN
Weighted
Authorized licensed use limited to: University of Canberra. Downloaded on June 10,2023 at 14:01:12 UTC from IEEE Xplore. Restrictions apply.
[7] G. Altan, Y. Kutlu, Y. Garbi, A. Ö. Pekmezci, and S. Nural,
"Multimedia Respiratory Database (RespiratoryDatabase@TR):
Auscultation Sounds and Chest X-rays," ArXiv, vol.
abs/2101.10946, 2021.
[8] M. Aykanat, Ö. Kılıç, B. Kurt, and S. Saryal, "Classification of
lung sounds using convolutional neural networks," EURASIP
Journal on Image and Video Processing, vol. 2017, no. 1, pp. 1-9,
2017.
[9] H. I. Hussein, H. I. Dino, R. J. Mstafa, and M. M. Hassan, "Person-
independent facial expression recognition based on the fusion of
HOG descriptor and cuttlefish algorithm," Multimedia Tools and
Applications, vol. 81, no. 8, pp. 11563-11586, 2022.
[10] S. Z. H. Naqvi and M. A. Choudhry, "An automated system for
classification of chronic obstructive pulmonary disease and
pneumonia patients using lung sound analysis," Sensors, vol. 20,
no. 22, p. 6512, 2020.
[11] Z. Tariq, S. K. Shah, and Y. Lee, "Lung Disease Classification
using Deep Convolutional Neural Network," 2019 IEEE
International Conference on Bioinformatics and Biomedicine
(BIBM), pp. 732-735, 2019.
[12] J. Saldanha, S. Chakraborty, S. Patil, K. Kotecha, S. Kumar, and
A. Nayyar, "Data augmentation using Variational Autoencoders
for improvement of respiratory disease classification," Plos one,
vol. 17, no. 8, p. e0266467, 2022.
[13] V. Basu and S. Rana, "Respiratory diseases recognition through
respiratory sound with the help of deep neural network," 2020 4th
International Conference on Computational Intelligence and
Networks (CINE), Kolkata, India, 2020, pp. 1-6, doi:
10.1109/CINE48825.2020.234388.
[14] S. I. Khan and R. B. Pachori, “Automated classification of lung
sound signals based on empirical mode decomposition,” Expert
Syst. Appl., vol. 184, p. 115456, Dec. 2021, doi:
10.1016/J.ESWA.2021.115456
[15] M. T. García-Ordás, J. A. Benítez-Andrades, I. García-Rodríguez,
C. Benavides, and H. Alaiz-Moretón, "Detecting Respiratory
Pathologies Using Convolutional Neural Networks and
Variational Autoencoders for Unbalancing Data," Sensors, vol.
20, no. 4, p. 1214, 2020.
[16] L. Brunese, F. Mercaldo, A. Reginelli, and A. Santone, "A Neural
Network-Based Method for Respiratory Sound Analysis and Lung
Disease Detection," Applied Sciences, vol. 12, no. 8, p. 3877,
2022.
[17] L. Fraiwan, O. Hassanin, M. Fraiwan, B. Khassawneh, A. M.
Ibnian, and M. Alkhodari, "Automatic identification of respiratory
diseases from stethoscopic lung sound signals using ensemble
classifiers," Biocybernetics and Biomedical Engineering, vol. 41,
no. 1, pp. 1-14, 2021.
[18] J. Acharya and A. Basu, "Deep neural network for respiratory
sound classification in wearable devices enabled by patient
specific model tuning," IEEE transactions on biomedical circuits
and systems, vol. 14, no. 3, pp. 535-544, 2020.
[19] M. U. Khan, S. J. T. J. o. E. E. Aziz, and C. Sciences, "A novel
pulse plethysmograph signal analysis method for identification of
myocardial infarction, dilated cardiomyopathy, and hypertension,"
vol. 29, no. 2, pp. 962-977, 2021.
[20] M. U. Khan et al., "Expert hypertension detection system featuring
pulse plethysmograph signals and hybrid feature selection and
reduction scheme," vol. 21, no. 1, p. 247, 2021.
[21] F. Amjad, A. Malik, M. Bilal, M. U. Khan, and S. Aziz, "Diagnosis
of Cardiac Disorders Featuring Pulse Plethysmograph Signals," in
2021 Mohammad Ali Jinnah University International Conference
on Computing (MAJICC), 2021, pp. 1-6: IEEE.
[22] M. U. Khan, S. Aziz, F. Amjad, and M. Mohsin, "Detection of
dilated cardiomyopathy using pulse plethysmographic signal
analysis," in 2019 22nd International multitopic conference
(INMIC), 2019, pp. 1-5: IEEE.
[23] M. U. Khan, S. Aziz, K. Iqtidar, R. J. B. S. P. Fernandez-Rojas,
and Control, "Computer-aided diagnosis system for cardiac
disorders using variational mode decomposition and novel cepstral
quinary patterns," vol. 81, p. 104509, 2023.
[24] S. Aziz, M. U. Khan, K. Iqtidar, S. Ali, A. N. Remete, and M. A.
J. E. S. Javid, "Pulse plethysmograph signal analysis method for
classification of heart diseases using novel local spectral ternary
patterns," p. e13011, 2022.
[25] Z. Kanji, "Classification of Auscultation Sounds Using a Smart
System," Massachusetts Institute of Technology, 2022.
[26] K. Iqtidar, U. Qamar, S. Aziz, M. U. J. C. i. B. Khan, and
Medicine, "Phonocardiogram signal analysis for classification of
Coronary Artery Diseases using MFCC and 1D adaptive local
ternary patterns," vol. 138, p. 104926, 2021.
[27] S. Aziz, M. U. Khan, M. Alhaisoni, T. Akram, and M. J. S. Altaf,
"Phonocardiogram signal processing for automatic diagnosis of
congenital heart disorders through fusion of temporal and cepstral
features," vol. 20, no. 13, p. 3790, 2020.
[28] M. U. Khan et al., "A two-stage classification model integrating
feature fusion for coronary artery disease detection and
classification," vol. 81, no. 10, pp. 13661-13690, 2022.
[29] U. Riaz, S. Aziz, M. Umar Khan, S. A. A. Zaidi, M. Ukasha, and
A. J. C. I. Rashid, "A novel embedded system design for the
detection and classification of cardiac disorders," vol. 37, no. 4,
pp. 1844-1864, 2021.
[30] S. Aziz, M. U. Khan, A. Rehman, Z. Tariq, and K. J. E. S. Iqtidar,
"Computer‐aided diagnosis of COVID‐19 disease from chest x‐
ray images integrating deep feature extraction," vol. 39, no. 5, p.
e12919, 2022.
[31] F. Bouaziz, D. Boutana, and H. Oulhadj, "Diagnostic of ECG
arrhythmia using wavelet analysis and K-nearest neighbor
algorithm," in 2018 International conference on applied smart
systems (ICASS), 2018, pp. 1-6: IEEE.
Authorized licensed use limited to: University of Canberra. Downloaded on June 10,2023 at 14:01:12 UTC from IEEE Xplore. Restrictions apply.